Machine Learning to Predict Executive Function in Adolescents with Repaired d-TGA, TOF, and Fontan Palliation

Abstract

Objective

to identify predictors of impaired executive function (EF) in adolescents after surgical repair of critical congenital heart disease (CHD).

Study design

We analyzed patient factors, medical and surgical history, and family social class from three single-center studies of adolescents with d-transposition of the great arteries (d-TGA), tetralogy of Fallot (TOF), and Fontan repair. Machine learning models were developed using recursive partitioning to predict an EF composite score based on five subtests (population mean 10, SD 3) of the Delis-Kaplan Executive Function System.

Results

The sample included 386 patients (139 d-TGA, 91 TOF, 156 Fontan) of age 15.1 ± 2.1 (mean ± SD) years and an EF composite score of 8.6 ± 2.4. Family social class emerged as the most important predictive factor. The lowest (worst) mean EF score (5.3) occurred in patients with low to medium social class (Hollingshead index < 56) with one or more neurologic events and a diagnosis of TOF. The highest (best) mean score (9.7) occurred in subjects with high social class (Hollingshead index ≥ 56) and shorter duration of deep hypothermic circulatory arrest. Other factors predicting lower EF scores included low birth weight and a higher number of catheterizations.

Conclusion

In regression tree modeling, family social class was the strongest predictor of EF in adolescents with critical CHD, even in the presence of medical risk factors. Additional predictors included CHD diagnosis, birth weight, neurologic events, and number of procedures. These data highlight the importance of social class in mitigating risks of executive dysfunction in CHD.

Executive function, encompassing an array of higher-order neurocognitive abilities, is critical for success in school, work, and emotional well-being.¹ Among children, adolescents, and adults with complex congenital heart disease (CHD), the problems of executive function impairment are nearly twice as high as those in the normative population, predisposing to lower future employability as well as more risk-taking behaviors such as addiction.^2–5 Although some studies report significant improvements in executive functioning following intervention,^6–8 overall the evidence for efficacy of such interventions is mixed.⁹

Classification and regression tree (CART) analysis is a form of machine learning or data mining in which a decision tree is created to predict continuous variables (regression tree analysis) or discrete variables (classification tree analysis) by recursive partitioning.¹⁰ Compared with complex regression methods, this nonparametric approach is easy to understand and translate into clinical practice. regression tree analyses can also expose interactions among predictor variables that are difficult to uncover in traditional multivariable regression. We aimed to construct a model using this machine learning technique to predict adverse executive function outcomes in adolescents with CHD using data that are readily available to clinicians in their routine clinical care.

METHODS

Our study was based on analyses of a dataset that combined three previously-published single-center cohort studies of neurodevelopmental outcome in adolescents with complex congenital heart disease.^11–13 We used patient, intraoperative, and post-operative variables that were collected in common across the three studies to predict executive function. These studies, as well as ensuing data analyses, were approved by our institutional review board.

Subjects aged 10–19 years included adolescents with d-transposition of the great arteries (d-TGA) who participated in the Boston Circulatory Arrest Study;¹¹ adolescents with tetralogy of Fallot (TOF);¹² and pre-teens and adolescents with single-ventricle cardiac anatomy who underwent the Fontan operation.¹³ Entry criteria for each patient cohort are described below.

d-TGA group:

Inclusion criteria were targeted age 16 years and diagnosis of d-TGA with or without ventricular septal defect with scheduled arterial switch operation by 3 months of age. The d-TGA surgeries were performed between 1988–1992. Exclusion criteria included birth weight < 2.5 kg, a recognizable syndrome of congenital anomalies, an associated extracardiac anomaly of greater than minor severity, previous cardiac surgery such as coarctation repair prior to the arterial switch operation, or associated cardiac anomalies requiring aortic arch reconstruction or additional open surgical procedures. Although the Boston Circulatory Arrest Study, from which the d-TGA group was drawn, excluded patients with more than minor extracardiac anomalies or recognizable genetic syndromes, d-TGA patients rarely have associated genetic syndromes.

TOF group:

Inclusion criteria were age 13–16 years, diagnosis of tetralogy of Fallot with or without pulmonary atresia, and an interval of at least 3 months between the last cardiac surgery and neurodevelopmental testing. The TOF surgeries were performed between 1988–1996. Exclusion criteria included patients with trisomy 21 or lack of reading fluency by the primary caregiver. Patients were assigned to the genetic/phenotypic syndrome group based on medical history of a genetic/phenotypic syndrome of multiple anomalies or a finding of a genetic disorder on genetic testing.

Fontan group:

Inclusion criteria were age 10 to 19 years, diagnosis of single ventricle lesion, and history of Fontan procedure. The Fontan surgeries were performed between 1990–2003. Exclusion criteria included lack of reading fluency by primary caregiver in English, foreign residence, or cardiac transplantation or cardiac surgery within 6 months of neurodevelopmental testing. Individuals underwent genetic evaluation that included examination by a clinical geneticist and chromosomal microarray testing.

The Delis-Kaplan Executive Function System (DKEFS) is a widely used battery of laboratory executive function tasks with expected mean score of 10 (SD 3) in the normal population, with a lower score suggesting greater executive dysfunction and a score ≤ 7 considered to be of clinical concern.¹⁴ Our primary outcome, as previously described, was an executive function composite score derived by averaging an adolescent’s scores on five DKEFS subtests:^11–13 mean letter fluency and category fluency trials of Verbal Fluency, primary combined measure on Design Fluency, combined conditions score on Sorting, total consecutively correct score on Word Context, and total achievement score on Tower. Of note, DKEFS scores are standardized by age. We also report the Behavior Rating Inventory of Executive Function (BRIEF)-Parent Global Score, based on a parent-completed questionnaire that is used to evaluate executive function in children.¹⁵ The expected mean T-score on the BRIEF for the normal population is 50 (SD 10), and a higher score suggests greater executive dysfunction. A BRIEF score ≥ 60 may be of potential clinical concern.

We considered sociodemographic factors, intraoperative bypass variables for first surgery, and medical history in our analyses. Sociodemographic factors included sex, race/ethnicity, and Hollingshead family social class scores.¹⁶ Due to small counts, for race/ethnicity we compared non-Hispanic White participants to all others (participants who were Black, Asian/Pacific Islander, multiracial, or White of Hispanic ethnicity). Because the Fontan group had a wider age range at neurodevelopmental testing compared with the d-TGA and TOF groups, we also included age at testing in analyses in the combined data set and those restricted to the Fontan group. Intraoperative bypass variables included durations of deep hypothermic circulatory arrest (DHCA), cardiopulmonary bypass, and total support time at first operation. Of note, the majority of patients with single ventricle lesions aside from hypoplastic left heart syndrome (HLHS) did not require cardiopulmonary bypass at their first operation. Due to this reason, intraoperative bypass variables were not considered in models for the Fontan group. Medical history variables included gestational age, birth weight, diagnosis group (d-TGA, TOF, Fontan), genetic abnormality, age at first operation, total number of operations and cardiac catheterizations until the time of adolescent neurodevelopmental evaluation, perioperative complications, and postoperative neurologic events. Perioperative complications included sustained arrhythmia, extracorporeal membrane oxygenation (ECMO), cardiac tamponade, open chest, infection, sepsis, organ failure (liver or kidney), reoperation/reexploration, postoperative ICU stay > 1 week, respiratory arrest, and reintubation. Postoperative neurologic events included stroke, seizure, choreoathetosis, and meningitis.

Linear regression was used to examine univariable associations between executive function outcomes and predictor variables. Stepwise regression based on univariable P < 0.20, an improvement in the Akaike Information Criterion, and multivariable P < 0.05 was conducted to determine independent predictors of executive function. In addition, we used regression trees⁹ as a machine learning technique to develop alternative prediction models which are easier to interpret clinically and readily allow for assessment of effect modification. Recursive partitioning with minimum terminal node size 10 and complexity parameter 0.02 was used to create pruned regression trees in the entire sample, as well as by diagnosis group. The proportion of variability explained (R²) and the root mean squared error (RMSE), or estimated standard deviation around the fitted values, were used to assess and compare the prediction models. To adjust for possible overfitting, the RMSE was re-estimated using 10-fold cross validation (CV). ^17–19 A 10-fold CV approach randomly splits the original sample into ten approximately equal sized groups (stratified by diagnosis group) and performs the fitting procedure a total of ten times, with each fit being performed on a training set consisting of 90% of the original sample and the remaining 10% used as a validation set. We then estimated RMSE 10 times based on the validation sets and used the average of these to calculate a less biased estimate of the standard deviation around the fitted values. To assess the stability of the predictors selected for our regression trees, we used a bootstrapping approach. ^17–19 We generated 1,000 bootstrapped samples (sampling with replacement from the original sample stratified by diagnosis group) and fit regression trees to each bootstrapped sample as described above. We then report which predictors occurred ≥ 10% of the time as either a first-level tree split or ≥ 25% of the time as a first- or second-level tree split. CART analyses were completed using the rpart and rpart.plot libraries in R version 3.6.0 (Foundation for Statistical Computing, Vienna, Austria).

RESULTS

The demographic and clinical characteristics of the combined cohort and the three diagnostic groups are summarized in Table I. In brief, the analytic sample included 386 patients (139 d-TGA, 91 TOF, 156 Fontan) of age 15.1 ± 2.1 (mean ± SD) years. Among the d-TGA group, 32 (23%) had a ventricular septal defect. The TOF group included 26 (29%) individuals with pulmonary atresia. Among the Fontan group, the first operation was open, e.g., using cardiopulmonary bypass, in 92 patients (59%), including 63 (40%) who had previously undergone a Norwood operation. The remaining 64 patients (41%) had a closed first operation.

Table I.

Demographic and clinical characteristics of participants

Variable	All Groups	d-TGA	TOF	Fontan
	N = 386	N = 139	N = 91	N = 156
Preoperative
Birth weight (kg)	3.3 (0.6)	3.6 (0.4)	3.1 (0.7)	3.3 (0.6)
Gestational age (wk)	39.3 (2.0)	39.8 (1.2)	39.2 (2.3)	38.9 (2.2)
Male sex	251 (65%)	106 (76%)	50 (55%)	95 (61%)
Race
White	362 (94%)	131 (94%)	86 (95%)	145 (93%)
Black	9 (2%)	2 (1%)	2 (2%)	5 (3%)
Asian/Pacific Islander	8 (2%)	3 (2%)	1 (1%)	4 (3%)
Multiracial	7 (2%)	3 (2%)	2 (2%)	2 (1%)
Hispanic ethnicity	32 (8%)	6 (4%)	7 (8%)	19 (12%)
Family social class*	48 (13)	46 (12)	49 (12)	50 (13)
Genetic diagnosis	39 (10%)	0 (0%)	23 (25%)†	16 (10%)‡
Operative
Age at first operation (d)	8 [4–54]	6 [4–9]	112 [43–200]	6 [3–21]
DHCA time (min)	33 (24)	36 (22)	26 (24)	35 (28)§
Bypass time (min)	94 (38)	108 (33)	74 (35)	88 (41)§
Total support time (min)	127 (41)	143 (32)	101 (36)	123 (47)§
Number of perioperative complications¶	1 [0–2]	1 [0–2]	0 [0–1]	1 [0–2]
Postoperative
Any neurologic event‖	94 (24%)	37 (27%)	18 (20%)	39 (25%)
Seizures	57 (15%)	23 (17%)	11 (12%)	23 (15%)
Number of catheterizations	2 [1–4]	1 [0–1]	2 [0–5]	4 [3–5]
Number of operations	2 [1–3]	1 [1–1]	2 [1–3]	3 [3–3]
Developmental testing
Age at testing (years)	15.1 (2.1)	16.1 (0.5)	14.6 (1.2)	14.5 (2.9)
DKEFS Composite**	8.6 (2.4)	9.0 (2.1)	8.2 (2.9)	8.6 (2.2)
DKEFS Composite ≤ 7	81 (21%)	21 (15%)	25 (28%)	35 (23%)
BRIEF-Parent Global††	56.7 (12.4)	54.8 (12.2)	58.0 (13.1)	57.6 (21.1)
BRIEF-Parent Global ≥ 60	150 (39%)	43 (31%)	38 (42%)	69 (45%)
Full-scale IQ	93.4 (18.1)	98.4 (14.9)	89.0 (22.8)	91.6 (16.8)
Full-scale IQ ≤ 70	42 (11%)	5 (4%)	19 (21%)	18 (12%)

Values are mean (SD), number (%), or median [IQR]. All variables missing <3% except as noted.

^*Score on Hollingshead Four Factor Index of Social Status (range 8–66), with higher scores indicating higher social status.

^†Previously known genetic/phenotypic syndromes included 22q11 (n=11), VATER/VACTERL (n=8), fetal alcohol syndrome (n=2), Alagille’s syndrome (n=1), and Turner’s syndrome (n=1).

^‡Previously known genetic/phenotypic syndromes included VACTERL (n=2), 22q11 (n=1), CHARGE syndrome (n=1), hyperphenylalaninemia (n=1), and Kartagener syndrome (n=1) and 10 subjects had a pathogenic variant on array comparative genomic hybridization analysis.

^§Data missing for 18/92 patients (20%) undergoing an open surgery and not pertinent for 64 patients undergoing a closed surgery.

^¶Includes perioperative arrhythmia, ECMO, tamponade, open chest, infection, sepsis, organ failure, reoperation/reexploration, postoperative ICU stay > 1 week, respiratory arrest, and reintubation.

^‖Includes stroke, seizure, choreoathetosis, and meningitis.

^**DKEFS Composite Score available for 381 subjects (138 d-TGA, 88 TOF, 155 Fontan).

^††BRIEF-Parent Global Score available for 382 subjects (138 d-TGA, 90 TOF, 154 Fontan).

In the combined cohort, the average EF composite score on the DKEFS was 8.6 ± 2.4, and 81 patients (21%) had an at-risk composite score of ≤ 7, considered to be of clinical concern. On the BRIEF-Parent Global Score, 150 (39%) individuals had scores ≥ 60, of potential clinical importance. The EF composite score differed significantly across the three diagnostic groups (P = 0.03) with the d-TGA group having the highest (best) scores. The BRIEF-Parent Global Score did not differ significantly across the three diagnostic groups (P = 0.09).

Using stepwise linear regression in the combined cohort excluding subjects with missing intraoperative bypass variables (DHCA time, bypass time, and total support time), independent predictors of lower (worse) EF composite scores included lower Hollingshead index of social class, presence of a genetic diagnosis, more perioperative complications, history of any postoperative neurologic event, and a greater number of cardiac operations (based on N = 284), accounting for 27% of the variability in EF composite scores with an RMSE of 2.1 and CV RMSE of 2.1 (Table II). When intraoperative bypass variables were excluded as candidate variables, independent predictors of lower EF composite scores continued to include lower social class, presence of a genetic diagnosis, more perioperative complications, history of any postoperative neurologic event, and greater number of operations, but added predictors included race/ethnicity (non-Hispanic White vs. all other participants) and TOF diagnosis relative to Fontan diagnosis (N = 371). This model accounted for 26% of the variability in EF composite scores with an RMSE of 2.0 and CV RMSE of 2.1. Independent predictors of higher (worse) BRIEF-Parent Global Scores included younger gestational age, lower social class, presence of a genetic diagnosis, and history of any postoperative neurologic event. All three 10-fold CV RMSE values were <5% larger than the original model RMSE estimates, suggesting that overfitting was not a concern.

Table II.

Stepwise linear regression results for DKEFS Composite and BRIEF-Parent Global scores

Outcome	Variable	Beta [95% CI]	P value
DKEFS Composite N = 284 R2 = 27% RMSE = 2.1 CV RMSE = 2.1	Family social class*	0.05 [0.03, 0.07]	<0.001
	Genetic diagnosis	−2.11 [−3.01, −1.21]	<0.001
	DHCA time (min)	−0.01 [−0.02, −0.00]	0.01
	Number of perioperative complications‡	−0.16 [−0.32, −0.01]	0.04
	Any neurologic event†	−1.41 [−1.99, −0.83]	<0.001
	Number of operations	−0.31 [−0.56, −0.06]	0.01
DKEFS Composite (Surgical support times excluded from candidacy) N = 371 R2 = 26% RMSE = 2.0 CV RMSE = 2.1	Non-Hispanic White	0.78 [0.14, 1.43]	0.01
	Family social class*	0.06 [0.04, 0.07]	<0.001
	Genetic diagnosis	−1.51 [−2.28, −0.73]	<0.001
	Number of perioperative complications‡	−0.24 [−0.38, −0.11]	<0.001
	Any neurologic event†	−1.09 [−1.59, −0.58]	<0.001
	Number of operations	−0.35 [−0.61, −0.10]	0.008
	d-TGA (vs. Fontan)	−0.29 [−0.98, 0.40]	0.46
	TOF (vs. Fontan)	−0.82 [−1.45, −0.20]	0.02
BRIEF-Parent Global N = 369 R2 = 9% RMSE = 11.7 CV RMSE = 12.2	Gestational age < 39 weeks	3.36 [0.61, 6.11]	0.02
	Family social class*	−0.18 [−0.28, −0.08]	<0.001
	Genetic diagnosis	5.11 [1.00, 9.22]	0.01
	Any neurologic event†	4.70 [1.89, 7.52]	<0.001

^*Score on Hollingshead Four Factor Index of Social Status (range 8–66), with higher scores indicating higher social status.

^†Includes stroke, seizure, choreoathetosis, and meningitis.

^‡Includes perioperative arrhythmia, ECMO, tamponade, open chest, infection, sepsis, organ failure, reoperation/reexploration, postoperative ICU stay > 1 week, respiratory arrest, and reintubation.

In regression tree analysis of EF composite scores in the combined cohort (Figure 1A), social class emerged as the strongest predictive factor, appearing as the first branch of the tree. The worst (lowest) mean EF composite score (5.3) was found in 15 patients with Hollingshead index of social class < 56 with one or more postoperative neurologic events and a diagnosis of TOF. The next lowest mean score (6.2) occurred in 17 patients with Hollingshead index < 56 with one or more postoperative neurologic events, a diagnosis of Fontan or d-TGA, and four or more cardiac catheterizations. The best mean score (9.7) occurred in 110 patients with a Hollingshead index ≥ 56 and shorter duration of DHCA, including those subjects with a missing DHCA time. This model explained 26% of the variability in EF composite scores with an RMSE of 2.1 and a 10-fold CV RMSE of 2.3 (an increase of 11%). Based on bootstrapped samples, first-level splits occurred most often with Hollingshead index (45%), postoperative neurologic event (17%), presence of a genetic diagnosis (17%), or duration of DHCA (12%). First- or second-level splits occurred most often with Hollingshead index (75%), presence of a genetic diagnosis (36%), postoperative neurologic event (35%), birth weight (30%), and duration of DHCA (30%). For EF composite scores, R² and RMSE values from the regression tree slightly underperformed compared with stepwise linear regression, but the regression tree is easier to interpret.

Figure 1.

Regression tree for the combined cohort for A, DKEFS Composite Score (R² = 26%, RMSE = 2.1, CV RMSE = 2.3). B, BRIEF-Parent Global Score (R² = 9%, RMSE = 11.8, CV RMSE = 13.0). Each node indicates the mean DKEFS composite score or BRIEF-Parent Global Score (top number) and sample size (bottom number) for patients with the characteristics along the paths above the node. Soc class = Hollingshead Four Factor Index of Social Status, Neuro = Neurologic, Caths = Catheterization.

Regression tree analysis of BRIEF-Parent Global Scores in the combined cohort (Figure 1B) showed higher family social class, higher gestational age, and absence of any postoperative neurologic event as predictive of lower (better) scores, indicating lower perception of executive dysfunction by parent report, explaining 9% of the variability in BRIEF-Parent Global Scores with an RMSE of 11.8 and a 10-fold CV RMSE of 13.0 (an increase of 10%). The worst (highest) mean BRIEF-Parent Global Score (59.3) was found in 173 patients with Hollingshead index of social class < 49 regardless of medical factors. The best mean score (50.3) occurred in 97 patients with a Hollingshead index ≥ 49, gestational age ≥ 40 weeks, and with no postoperative neurologic event. Based on bootstrapped samples, first-level splits occurred most often with Hollingshead index (43%), birth weight (13%), or postoperative neurologic event (13%). Similarly, first- or second-level splits occurred most often with Hollingshead index (73%), birth weight (37%), and postoperative neurologic event (26%). For BRIEF-Parent Global Scores, the R² and RMSE values from the regression tree were comparable with those from stepwise linear regression.

We also performed regression tree analysis for EF composite scores in each diagnostic group. Among participants with d-TGA (Figure 2A), family social class emerged as the most predictive variable in the decision tree (R² = 26%, RMSE = 1.9, 10-fold CV RMSE = 2.5) and it was the first-level split in 66% of bootstrapped trees. Among the participants with TOF (Figure 2B), postoperative neurologic events emerged as the most predictive variable (R² = 46%, RMSE = 2.1, 10-fold CV RMSE = 2.5). Postoperative neurologic event was the first-level split in 56% of bootstrapped trees, followed by genetic diagnosis (11%) and Hollingshead index (10%). In the Fontan group (Figure 2C), perioperative complications emerged as the most predictive variable with Hollingshead index splitting at a lower level in the tree (R² = 26%, RMSE = 1.9, 10-fold CV RMSE = 2.5). Hollingshead index was the most common first-level split in 40% of bootstrapped trees, followed by perioperative complications (17%) and birth weight (11%).

Figure 2.

Diagnosis-specific regression trees for the DKEFS Composite Score. A, d-TGA (R² = 25%, RMSE = 1.8, CV RMSE = 2.1). B, TOF (R² = 46%, RMSE = 2.1, CV RMSE = 2.5). C, Fontan, R² = 24%, RMSE = 1.9, CV RMSE = 2.5). Each node indicates the mean DKEFS composite score (top number) and sample size (bottom number) for patients with the characteristics along the paths above the node. Soc class = Hollingshead Four Factor Index of Social Status, DHCA = Deep hypothermic circulatory arrest, Neuro = Neurologic.

DISCUSSION

Among the array of neuropsychological domains impacted by congenital heart disease, perhaps the most profound deficits occur in executive function.²⁰ In this study, we developed a machine learning model including clinically available patient data that can be used by clinicians to identify children with complex CHD who have a higher risk of developing executive function deficits. A major feature of all models is the strong contribution of family social class to executive function. Family social class is known to be correlated with a broad array of cardiovascular outcomes, as well as with neurodevelopment outcomes in CHD.^21,22 We show that the best EF scores were predicted in patients with higher family social class and shorter duration of DHCA. Additional predictors included DHCA time, CHD diagnosis, birth weight, neurologic events, and number of procedures.

Predictors of deficits in executive function in children with major congenital heart disease have previously been demonstrated using standard regression analysis. These have included birth weight, gestational age, race, cardiopulmonary bypass support times, postoperative seizures, and postoperative complications.^{11–13,20,23,24} Our current study using a machine learning algorithm builds upon these earlier findings by exposing the interaction of socioeconomic status with patient characteristic and medical risk factors.

Socioeconomic disparities in childhood executive function in the general population have been well studied. Lower socioeconomic status (SES) has been strongly correlated with poor EF; indeed socioeconomic disparities are more common in executive function than in other neurocognitive domains.^25–29 The pathways by which low parental SES might impact childhood executive function include familial psychological stress and low cognitive stimulation.³⁰ Additional pathways include poorer nutrition and schooling, greater exposure to environmental toxicants, and reduced access to medical care.^31–35 Such adversities in crucial developmental periods during childhood have been postulated to contribute to adult disease.³⁶

In our study in patients with CHD, high family social class seemed to mitigate EF deficits even in the presence of medical risk factors. This finding is consistent with work of Benavente-Fernández et al, in which higher maternal education modified the relationship between early life brain injury and later cognitive outcome in very preterm children.³⁷ Similarly, deleterious effects of in utero and early life exposure to air pollution and lead on child development have been shown to be enhanced in the setting of socioeconomic disadvantage.^38,39 The role of social class in attenuating deleterious effects of CHD on executive function suggests potential avenues for intervention.

The neural underpinnings of the interaction between social class and vulnerability to medical risk factors for executive dysfunction in CHD are uncertain. In the normal population, socioeconomic disparities have been associated with brain structure and function,⁴⁰ and the mechanisms of this relationship are an active area of investigation. For example, the dorsolateral prefrontal cortex volume has been found to be a mediator between socioeconomic status and executive function in white but not black adults.⁴¹ Brito et al found that higher family income lessened the association of greater cortical thickness with worse executive function scores.⁴² In CHD patients, studies of brain MRI and cognition have shown associations of abnormalities of brain volume, white matter injury, and network topology with cognitive function, including executive function. Future studies should explore whether socioeconomic factors serve as effect modifiers of the relationship between these CNS abnormalities in patients with CHD and executive function.^43–49

There are several limitations of this study. Participants were studied at a single center, potentially limiting the generalizability of our findings. The three studies from which our participant data were derived were conducted in different time periods, over which surgical and perfusion techniques evolved. Many risk factors in our models are interrelated, limiting causal inference. The small number of patients at terminal nodes limits outcome prediction in some cases. Our models analyzed a composite executive function score derived from the DKEFS, which combines a number of different component skills together. It is possible that the risk factors for lower performance differ among the component skills, and that separate analyses of the components would have identified different prediction models. Participants with d-TGA had been enrolled in infancy in the Boston Circulatory Arrest Study, in which exclusion criteria included birth weight <2.5 kg and known genetic syndromes; however, individuals with simple d-TGA have a normal distribution of birth weight and rarely are affected by genetic syndromes. We did not consider IQ as a predictor variable because it was measured concurrently with executive function. Most importantly, we aimed to construct our models using patient information readily available to a clinical provider. Data on intraoperative cardiopulmonary bypass times were missing for participants in the Fontan group who had closed first operations or who had their first open surgery at other institutions. Finally, we used the Hollingshead Four Factor Index to assess family socioeconomic status. Although this index is older and has come under criticism as unvalidated,⁵⁰ it is nonetheless the most widely used measure in studies of cardiac neurodevelopment.

Using machine learning models, family social class emerged as the strongest predictive factor for EF in adolescents with complex CHD. Additional predictive factors included CHD diagnosis and medical factors including lower birth weight, genetic diagnosis, neurologic events, DHCA time, perioperative complications, and more cardiac catheterizations. If validated in multi-center studies, these results could provide guidance for targeting therapeutic intervention in children with complex CHD.

Abbreviations:

BRIEF

Behavior Rating Inventory of Executive Function

CHD

Congenital heart disease

CART

Classification and regression tree

CV

Cross validation

DKEFS

Delis-Kaplan Executive Function System

d-TGA

d-tranposition of the great arteries

DHCA

Deep hypothermic circulatory arrest

EF

Executive function

HLHS

Hypoplastic left heart syndrome

RMSE

Root mean squared error

SES

Socioeconomic status

TOF

Tetralogy of Fallot